American Water

Attend the Innovation Revolution - Nov 16th

Disclaimer

• This is a long presentation, but I can take questions during. Just raise your hand high.
• I’m speaking for myself this evening. All of my opinions are my own and do not necessarily reflect those of RStudio, Inc.

Get The Things You Need

devtools::install_github('rstudio/tensorflow')
devtools::install_github('rstudio/keras')
devtools::install_github('rstudio/tfdeploy')
devtools::install_github('rstudio/cloudml')
devtools::install_github('rstudio/rsconnect')
# Run Python inside R - surprisingly fast! 
# Also translates numpy dataframes to R dataframes on the fly 😮
devtools::install_github('rstudio/reticulate') 

GPU Acceleration

• nVidia GPU
• Proprietary drivers
• CUDA Toolkit 9.0
• cuDNN SDK 7.2+

https://www.tensorflow.org/install/gpu

More on GPU Acceleration

• Multiple GPUs are supported
• Amazon EC2 P3
• Google Cloud Compute GPU
• Microsoft Azure NC Series

Hosting Side Note

• Google Cloud ML: https://cloud.google.com/ml-engine/
• RStudio Connect

My tools

• RStudio (Sorry, shameless plug, I know.)
  • You’ll want the latest versions for Jobs support!
  • Comes in Desktop & Server, FOSS and Pro flavors

What can you do with TensorFlow?

library(tensorflow)
tensorflow::tf_version()

## [1] '1.9'

Use MNIST Data

sess <- tf$Session()
data <- tf$
  contrib$
  learn$
  datasets$
  mnist$
  read_data_sets('MNIST', one_hot = T)
len <- length(data$test$images[1,])[[1]]
len

## [1] 784

and…

input <- tf$placeholder(
  dtype = tf$float32,
  shape = shape(NULL, len)
  )
# y = WX + b
W <- tf$Variable(tf$zeros(shape(len, 10)))
B <- tf$Variable(tf$zeros(shape(10)))
Y <- tf$nn$softmax(tf$matmul(input, W) + B)

and…

# Output
Y_ <- tf$placeholder(tf$float32, shape(NULL, 10L))

# Cross-entropy loss function
cross_entropy <- tf$reduce_mean(-tf$reduce_sum(Y_ * tf$log(Y), reduction_indices=1L))

train_step <- tf$train$GradientDescentOptimizer(0.5)$minimize(cross_entropy)

sess$run(tf$global_variables_initializer())

Training

for (i in 1:2000) {
  batches <- data$train$next_batch(100L)
  batch_xs <- batches[[1]]
  batch_ys <- batches[[2]]
  sess$run(train_step,
           feed_dict = dict(input = batch_xs, Y_ = batch_ys))
}

Accuracy

correct_prediction <- tf$equal(tf$argmax(Y, 1L), tf$argmax(Y_, 1L))
accuracy <- tf$reduce_mean(tf$cast(correct_prediction, tf$float32))

sess$run(accuracy, feed_dict=dict(input = data$test$images, Y_ = data$test$labels))

## [1] 0.9179

Basic Import

library(keras)
keras::is_keras_available()

## [1] TRUE

Same Idea

library(zeallot)

# load data
c(c(x_train, y_train), c(x_test, y_test)) %<-% dataset_mnist()
# reshape and rescale
x_train <- array_reshape(x_train, dim = c(nrow(x_train), 784)) / 255
x_test <- array_reshape(x_test, dim = c(nrow(x_test), 784)) / 255
# one-hot encode response
y_train <- to_categorical(y_train, 10)
y_test <- to_categorical(y_test, 10)

Model is a bit more declarative

# define and compile model
model <- keras_model_sequential()
model %>%
  layer_dense(units = 32, activation = 'relu', input_shape = c(784),
              name = "image") %>%
  layer_dense(units = 16, activation = 'relu') %>%
  layer_dense(units = 10, activation = 'softmax',
              name = "prediction") %>%
  compile(
    loss = 'categorical_crossentropy',
    optimizer = optimizer_rmsprop(),
    metrics = c('accuracy')
  )

Training

# train model
history <- model %>% fit(
  x_train, y_train,
  epochs = 30, batch_size = 128,
  validation_split = 0.2
)

Step 1: Get a Script

How About That TensorFlow MNIST Script?

Step 2: Execute that Job

Why do I want this?

• Actually decent models require tons of data and lots of training time
• Queue a job on a server and go do more important things
• Do more than one at once

What kind of data do I want?

• Predictable/Classifiable data
• LOTS: Deep learning in particular needs a high N
• Normalized? Not necessarily- deep learning techniques are resilient to poor normalization, provided you have enough.

Where?

• https://www.kaggle.com/datasets
• https://www.opendataphilly.org/
• https://www.data.gov/

Tensorflow

export_savedmodel(
  sess,
  "tensorflow-mnist",
  inputs = list(images = input),
  outputs = list(scores = Y))

Keras

export_savedmodel(model, "keras-mnist")

TFDeploy

library(tfdeploy)
serve_savedmodel('tensorflow-mnist', browse=T)

CloudML

library(cloudml)
cloudml_deploy('tensorflow-mnist', name='tensorflow_mnist', version='tensorflow_mnist_1')

RStudio Connect

library(rsconnect)
deployTFModel('tensorflow-mnist', account = 'me', server = 'example.com')

RStudio Connect Tensorflow Demo

Enjoy the show

Mobile devices

https://www.tensorflow.org/lite/

1. Convert your saved model using https://www.tensorflow.org/lite/convert/
2. Include in mobile projects
3. ???
4. Profit! (No, really…)

Web Browsers

• Yes, really.
• https://js.tensorflow.org/
• Surprisingly high performance